Google Launches LiteRT.js for High-Performance AI Inference in Web Browsers

Google has announced LiteRT.js, a JavaScript library that brings its on-device LiteRT inference runtime directly to web browsers via WebAssembly. The library supports native hardware acceleration across CPU (via XNNPACK), GPU (via ML Drift and WebGPU), and NPU (via the experimental WebNN API), and enables one-step conversion of PyTorch models through LiteRT Torch. Google reports up to 3x performance improvements over existing web AI runtimes such as TensorFlow.js across tasks including object detection, audio processing, and text generation.

LiteRT.js is positioned as an evolution from TensorFlow.js for developers working with .tflite models, sharing a unified cross-platform stack with LiteRT on Android, iOS, and desktop. By running inference entirely client-side, it eliminates server costs and reduces latency while keeping user data local. An initial npm package and a set of demos — including a browser-based vector search powered by EmbeddingGemma — are available now.